path: root/lib/Analysis/DataStructure/EquivClassGraphs.cpp

//===- EquivClassGraphs.cpp - Merge equiv-class graphs & inline bottom-up -===//
// 
//                     The LLVM Compiler Infrastructure
//
// This file was developed by the LLVM research group and is distributed under
// the University of Illinois Open Source License. See LICENSE.TXT for details.
// 
//===----------------------------------------------------------------------===//
//
// This pass is the same as the complete bottom-up graphs, but
// with functions partitioned into equivalence classes and a single merged
// DS graph for all functions in an equivalence class.  After this merging,
// graphs are inlined bottom-up on the SCCs of the final (CBU) call graph.
//
//===----------------------------------------------------------------------===//

#define DEBUG_TYPE "ECGraphs"
#include "EquivClassGraphs.h"
#include "llvm/Module.h"
#include "llvm/Pass.h"
#include "llvm/Analysis/DataStructure/DSGraph.h"
#include "llvm/Analysis/DataStructure/DataStructure.h"
#include "llvm/Support/CallSite.h"
#include "Support/Debug.h"
#include "Support/SCCIterator.h"
#include "Support/Statistic.h"
#include "Support/EquivalenceClasses.h"
#include "Support/STLExtras.h"
using namespace llvm;

namespace llvm {
  namespace PA {
    Statistic<> NumFoldGraphInlines("Inline equiv-class graphs bottom up",
                                    "Number of graphs inlined");
    
  } // End PA namespace
} // End llvm namespace


namespace {
  RegisterAnalysis<llvm::PA::EquivClassGraphs> X("equivdatastructure",
                    "Equivalence-class Bottom-up Data Structure Analysis");
  Statistic<> NumEquivBUInlines("equivdatastructures", "Number of graphs inlined");
}


// getDSGraphForCallSite - Return the common data structure graph for
// callees at the specified call site.
// 
Function *llvm::PA::EquivClassGraphs::getSomeCalleeForCallSite(const CallSite &CS) const {
  Function *thisFunc = CS.getCaller();
  assert(thisFunc && "getDSGraphForCallSite(): Not a valid call site?");
  DSNode *calleeNode = CBU->getDSGraph(*thisFunc).
    getNodeForValue(CS.getCalledValue()).getNode();
  std::map<DSNode*, Function *>::const_iterator I =
    OneCalledFunction.find(calleeNode);
  return (I == OneCalledFunction.end())? NULL : I->second;
}

// computeFoldedGraphs - Calculate the bottom up data structure
// graphs for each function in the program.
//
void llvm::PA::EquivClassGraphs::computeFoldedGraphs(Module &M) {
  CBU = &getAnalysis<CompleteBUDataStructures>();

  // Find equivalence classes of functions called from common call sites.
  // Fold the CBU graphs for all functions in an equivalence class.
  buildIndirectFunctionSets(M);

  // Stack of functions used for Tarjan's SCC-finding algorithm.
  std::vector<Function*> Stack;
  hash_map<Function*, unsigned> ValMap;
  unsigned NextID = 1;

  if (Function *Main = M.getMainFunction()) {
    if (!Main->isExternal())
      processSCC(getOrCreateGraph(*Main), *Main, Stack, NextID, ValMap);
  } else {
    std::cerr << "Fold Graphs: No 'main' function found!\n";
  }
  
  for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I)
    if (!I->isExternal() && !FoldedGraphsMap.count(I))
      processSCC(getOrCreateGraph(*I), *I, Stack, NextID, ValMap);

  getGlobalsGraph().removeTriviallyDeadNodes();
}


// buildIndirectFunctionSets - Iterate over the module looking for indirect
// calls to functions.  If a call site can invoke any functions [F1, F2... FN],
// unify the N functions together in the FuncECs set.
//
void llvm::PA::EquivClassGraphs::buildIndirectFunctionSets(Module &M) {
  const ActualCalleesTy& AC = CBU->getActualCallees();

  // Loop over all of the indirect calls in the program.  If a call site can
  // call multiple different functions, we need to unify all of the callees into
  // the same equivalence class.
  Instruction *LastInst = 0;
  Function *FirstFunc = 0;
  for (ActualCalleesTy::const_iterator I=AC.begin(), E=AC.end(); I != E; ++I) {
    if (I->second->isExternal())
      continue;                         // Ignore functions we cannot modify

    CallSite CS = CallSite::get(I->first);

    if (CS.getCalledFunction()) {       // Direct call:
      FuncECs.addElement(I->second);    // -- Make sure function has equiv class
      FirstFunc = I->second;            // -- First callee at this site
    } else {                            // Else indirect call
      // DEBUG(std::cerr << "CALLEE: " << I->second->getName()
      //       << " from : " << I->first);
      if (I->first != LastInst) {
        // This is the first callee from this call site.
        LastInst = I->first;
        FirstFunc = I->second;
	// Instead of storing the lastInst For Indirection call Sites we store
	// the DSNode for the function ptr arguemnt
	Function *thisFunc = LastInst->getParent()->getParent();
	DSNode *calleeNode = CBU->getDSGraph(*thisFunc).getNodeForValue(CS.getCalledValue()).getNode();
        OneCalledFunction[calleeNode] = FirstFunc;
        FuncECs.addElement(I->second);
      } else {
        // This is not the first possible callee from a particular call site.
        // Union the callee in with the other functions.
        FuncECs.unionSetsWith(FirstFunc, I->second);
#ifndef NDEBUG
	Function *thisFunc = LastInst->getParent()->getParent();
	DSNode *calleeNode = CBU->getDSGraph(*thisFunc).getNodeForValue(CS.getCalledValue()).getNode();
        assert(OneCalledFunction.count(calleeNode) > 0 && "Missed a call?");
#endif
      }
    }

    // Now include all functions that share a graph with any function in the
    // equivalence class.  More precisely, if F is in the class, and G(F) is
    // its graph, then we include all other functions that are also in G(F).
    // Currently, that is just the functions in the same call-graph-SCC as F.
    // 
    DSGraph& funcDSGraph = CBU->getDSGraph(*I->second);
    const DSGraph::ReturnNodesTy &RetNodes = funcDSGraph.getReturnNodes();
    for (DSGraph::ReturnNodesTy::const_iterator RI=RetNodes.begin(),
           RE=RetNodes.end(); RI != RE; ++RI)
      FuncECs.unionSetsWith(FirstFunc, RI->first);
  }

  // Now that all of the equivalences have been built, merge the graphs for
  // each equivalence class.
  //
  std::set<Function*> &leaderSet = FuncECs.getLeaderSet();
  DEBUG(std::cerr << "\nIndirect Function Equivalence Sets:\n");
  for (std::set<Function*>::iterator LI = leaderSet.begin(),
	 LE = leaderSet.end(); LI != LE; ++LI) {

    Function* LF = *LI;
    const std::set<Function*>& EqClass = FuncECs.getEqClass(LF);

#ifndef NDEBUG
    if (EqClass.size() > 1) {
      DEBUG(std::cerr <<"  Equivalence set for leader " <<LF->getName()<<" = ");
      for (std::set<Function*>::const_iterator EqI = EqClass.begin(),
             EqEnd = EqClass.end(); EqI != EqEnd; ++EqI)
        DEBUG(std::cerr << " " << (*EqI)->getName() << ",");
      DEBUG(std::cerr << "\n");
    }
#endif

    if (EqClass.size() > 1) {
      // This equiv class has multiple functions: merge their graphs.
      // First, clone the CBU graph for the leader and make it the 
      // common graph for the equivalence graph.
      DSGraph* mergedG = cloneGraph(*LF);

      // Record the argument nodes for use in merging later below
      EquivClassGraphArgsInfo& GraphInfo = getECGraphInfo(mergedG);
      for (Function::aiterator AI1 = LF->abegin(); AI1 != LF->aend(); ++AI1)
        GraphInfo.argNodes.push_back(mergedG->getNodeForValue(AI1));
      
      // Merge in the graphs of all other functions in this equiv. class.
      // Note that two or more functions may have the same graph, and it
      // only needs to be merged in once.  Use a set to find repetitions.
      std::set<DSGraph*> GraphsMerged;
      for (std::set<Function*>::const_iterator EqI = EqClass.begin(),
             EqEnd = EqClass.end(); EqI != EqEnd; ++EqI) {
        Function* F = *EqI;
        DSGraph*& FG = FoldedGraphsMap[F];
        if (F == LF || FG == mergedG)
          continue;
        
        // Record the "folded" graph for the function.
        FG = mergedG;
        
        // Clone this member of the equivalence class into mergedG
        DSGraph* CBUGraph = &CBU->getDSGraph(*F); 
        if (GraphsMerged.count(CBUGraph) > 0)
          continue;
        
        GraphsMerged.insert(CBUGraph);
        DSGraph::NodeMapTy